TavernAI ChatGPT (GPT3.5 Turbo) mod
Installation:
- Setup TavernAI with OpenAI. See https://rentry.org/spoonfed or https://rentry.org/moosetavernal, should be easy enough.
- Download https://anonfiles.com/hdk2W8a1z3/TavernAI_Mod_zip and unzip it into your TavernAI's folder, replacing files.
diff1.diff and diff2.diff are only useful to cooders, since everyone's modding TavernAI and can't get an idea of the changes in one mod or another, so I decided to at least share the diffs.
This mod is based on the OpenAI mod.
Known bugs:
- No support for personality (the one in advanced settings) or anchors (the code is a real mess), not sure if I'll try fixing it today or tomorrow
- Max token size is incorrect because TavernAI uses an old tokenizer incompatible with the Turbo model. To (mostly) fix it, go to index.html, search for "4096", and replace 350 with 500 (so it's like 4096 - 500).
Some info:
- Example messages are parsed by splitting by newline and searching for "{{user}}:" and "{{char}}:" where those placeholders are replaced with your current name and the char's name. If your example messages don't use the {{user}} and {{char}} template names (e.g. if they use "You:" and "Charname:" directly), the output will be worse!
Changelog (if I decide to actually fix something):
- Ver 0.21 - changed max context size to be 4096-350 to account for the completion and some extra just in case. Thanks to anon
- Ver 0.2 - 03/02/2023 13:20 UTC, fixed content length trimming (supposedly), changed default context length from 2048 to 4096 (both davinci and gpt35 support that). Thanks to anon
- Ver 0.1 - 03/02/2023 11:40 UTC, initial "release"
Old bugs (fixed):
- Fixed in 0.21 - Max context length of 4096 meant that if you had a context bigger than 4096-300 you won't get a new completion.
- Fixed in 0.2 - "Doesn't respect the max context length, so chats longer than 4k tokens will error for now. Will try to fix it soon."
Old versions:
Peace Through Power!
By CncAnon